.png){kind=icon}
Return
DANFlow: Depth-wise Separable Convolution and Attention-based Normalizing Flow
.png){kind=icon}
PRE
AI
DOI：10.3837/tiis.2026.04.019.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
While Generative AI (GenAI) offers enormous potential for anomaly localization, its transition to safety-critical industrial environments is often limited by reasoning instability, susceptibility to noise, and high computational demands. This study addresses these deployment bottlenecks by proposing a novel Normalizing Flows architecture called Depthwise separable convolution and Attention-based Normalizing Flow (DANFlow). Unlike existing generative models that struggle with the stringent latency requirements of high-speed production lines, DANFlow is specifically designed to localize industrial anomalies in realtime. Our architecture introduces the Extended Inverted Residual Bottleneck (EIRB), a lightweight module that enhances feature extraction while maintaining computational efficiency. Additionally, an integrated attention mechanism improves model robustness by prioritizing critical defect details and suppressing environmental noise. We also provide a comprehensive analysis of backbone feature selection, offering a practical framework for optimizing performance. Extensive experimentation has shown that DANFlow achieves stateof-the-art results across four industrial benchmark datasets (MVTecAD, BTAD, KolektorSDD2, VisA) in terms of AUROC and AUPRO. Notably, on MVTecAD, DANFlow achieves 99.09% AUROC and 98.51% AUPRO, significantly outperforms existing methods in inference speed (FPS) and parameter efficiency, making it highly suitable for industrial setting of anomaly localization. The source code is publicly available at https://github.com/nguyendothanhtruc/DANFlow.
Keywords:
Defect Localization
Generative AI
Manufacturing
Normalizing Flows
One-class Classification
Journal
IF:
0.9
Papers:
240
Citations:
1.7K